Flash-based Computing-in-memory Architectures with High-accuracy and Robust Reliabilities for General-purpose Applications

Flash memory is a promising candidate for novel computing-in-memory (CIM) architectures due to its mature technologies, including robust reliabilities, large capacities, high bit densities, and low cost. However, to design CIM-compatible operation schemes for flash cells, the traditional operations adopted for memory or storage usages should be modified and optimized. In this work, we review recent progress on flash-based CIM architectures, including the operation strategies to enhance the reliabilities, and its applications in high-precision calculations. On the one side, balancing the memory window can achieve 3~4 orders higher endurance (>10
⁸
), which can be utilized in on-chip AI training; on the other side, 3~4 bits/cell schemes can be used to perform large-scale matrix operations with ultra-low power consumption. All these indicate that flash memory has great potential to provide general-purpose CIM solutions in various applications that require high performance and low power.